I Spoke an Idea Into My Phone.
It Became a GitHub PR.

I spoke into my phone.

Later, those words became a GitHub Issue, then a GitHub PR.

At the start, I did not even have a brief. I had a few threads coming together: online discussions about ADHD, people's experiences of struggling to start important tasks, a public conversation about mentally warming up before difficult work, a BBC Worklife article about why tiny tasks can feel so difficult, and the role Flow Keyboard might play. The BBC article describes how unfamiliar or emotionally loaded tasks can grow disproportionately large in our minds^[4].

It was a mix of observations, guesses, and facts that needed checking. I opened Flow Keyboard and talked it all through. I had no outline, and I did not try to write a perfect prompt first.

Flow Keyboard turned that free-form dictation into text. I sent it to ChatGPT, and the idea started moving forward.

You do not need a finished thought to get started

New ideas rarely arrive fully formed. We sense that something is worth exploring before we can define the central question, the limits of what we know, or the final deliverable.

Demanding a rigorous brief at that point can leave us staring at a blank page. Voice offers an easier way in. In a smartphone survey study with 1,001 participants, published in 2024, Höhne, Gavras, and Claassen found that participants randomly assigned to answer open-ended questions by voice used more words and covered more topics^[1].

The experiment does not prove that speaking produces better ideas. It does suggest that making expression easier may encourage people to share more background and detail. For a fuller discussion of voice, working memory, and AI context, read Give AI Too Little Context, and It Has to Guess.

Spontaneous speech includes pauses, repetition, corrections, and crucial background added halfway through. The CHI 2024 Rambler study starts with this problem: dictation is efficient, but spontaneous speech is often wordy and disjointed. Extracting the gist and revising the overall structure can make it easier to use as written material^[2].

At this stage, Flow Keyboard captures thinking that is still taking shape. It accurately transcribes extended speech and handles sentence breaks, repetition, and structure, turning a loose idea into material I can read, check, and discuss.

1. 01

   I talk through the idea Flow Keyboard

   “Could we connect ADHD, difficulty starting tasks, the idea of mentally warming up before difficult work, and Flow Keyboard in a genuinely useful article?”

   No outline: observations, experiences, and information still to be verified are all mixed together.

2. 02

   ChatGPT assesses and researches the idea ChatGPT

   “This is worth exploring, but we should avoid medical claims. A more precise audience is people who struggle to focus and get started.”

   Research continues into the boundaries of ADHD, procrastination studies, the source of the warm-up account, and what the BBC article actually supports.

3. 03

   I add requirements for the next stage Flow Keyboard

   “Keep the research citations, then create a GitHub Issue so the work can continue.”

4. 04

   ChatGPT turns the discussion into a task GitHub

   The research questions, limits of the evidence, article structure, platform deliverables, and acceptance criteria go into one GitHub Issue.

The real value emerges over several rounds

I first asked ChatGPT to assess whether the idea held up.

The first round quickly exposed problems. “Flow Keyboard helps people with ADHD” risked implying a medical benefit. “Everyone has ADHD” blurred clinical diagnosis, attention difficulties, and everyday procrastination. We narrowed the audience to “people who struggle to focus and find it hard to start tasks.”

ChatGPT then looked into definitions of adult ADHD, procrastination research, the public account of mentally warming up before difficult work, the BBC article's discussion of unfamiliar or emotionally loaded tasks, and how voice input affects expression. Each piece of evidence changed the original idea a little: some judgments held up, some claims became narrower, and some gaps needed more questions.

I used Flow Keyboard to add requirements: keep the research citations, spell out the medical claims we must avoid, and turn the discussion into a GitHub Issue for the next stage.

By then, we had a body of context shaped by questions, fact-checking, and editorial decisions. Anyone taking over, human or AI, could understand why the idea was worth pursuing and see where the evidence still did not support a conclusion.

Research and create a content series on easing into tasks and difficulty getting started GitHub Issue

Central idea: Flow Keyboard turns the process of getting started into usable work

1. 01

   Research

   Establish the facts and their limits

   ADHD versus everyday procrastination · The source of the warm-up account · What the BBC article supports · The voice-response experiment and its limitations

2. 02

   Content creation

   Build complete content from the same evidence

   Research notes and a long-form Zhihu article · A visual storyboard for Xiaohongshu (RedNote) · Copy for Weibo and Jike

3. 03

   Execution

   Make the work ready to check, revise, and publish

   Render images and run automated checks · Submit a GitHub PR · Publish after human review

Final deliverables research.md · longform.md · 9 images · 4 platform versions · GitHub PR

A GitHub Issue turns discussion into execution

Once the direction was clear, ChatGPT organized the discussion into a GitHub Issue.

It contained the research questions, existing evidence, factual boundaries, content structure, platform deliverables, and acceptance criteria. That first rambling dictation had become a task another person or AI could pick up and continue.

Codex then read the GitHub Issue, repository guidelines, and previous content. It continued checking sources and produced research notes, a long-form Zhihu article, an illustrated Xiaohongshu post, a visual plan, and versions for multiple platforms. All changes went into a GitHub PR. A person then checked whether the story held together, the citations supported the claims, the product descriptions were accurate, and the image layouts had any overflow.

Flow Keyboard captures the thought. ChatGPT fills in the questions. GitHub preserves the boundaries. Codex carries out the work. A person makes the final judgment.

OpenAI positions ChatGPT Work for multistep work such as research, analysis, and documents, and Codex for code, tests, commands, reviews, and repository work^[3]. An idea captured on a phone can therefore move into different working environments depending on the task: development work goes to GitHub and Codex, while research and office work can become reports, spreadsheets, presentations, or other deliverables.

For a more specific example of explaining a complex development request to Codex, see Voice Input for Codex.

The time saved comes from keeping the context

Before, I would hold an idea in my head, reconstruct it when I got back to my computer, explain it again in ChatGPT, and then reorganize the conversation into a task when it was time to act.

Now the same context can keep moving: spoken, questioned, checked, turned into a task, and carried through to execution. My phone can be a starting point for work as well as a place to receive messages.

The workflow comes down to four steps:

1. Talk it through for two minutes with Flow Keyboard: What have I noticed? Why is it worth doing? What is still unclear?
2. Give the organized text to ChatGPT. Ask it to distinguish facts, assumptions, and judgments, and identify the questions that still need answers.
3. Turn development work into a GitHub Issue. For research and office work, define the goal, sources, outputs, and boundaries.
4. Let Codex or ChatGPT Work carry out the task, then have a person review the result and decide whether to use it.

AI can generate plenty of answers. The greater value is helping an idea that might otherwise disappear become something we can review, revise, and act on.

Start by saying it out loud.